Sign in Subscribe
Century Report

OpenAI Arms Open-Source Defenders - TCR 06/23/26

Five nations warn AI cyberattack power is months away, while the same week OpenAI hands machine-speed patching to open-source defenders.

Ben Linford

Ben Linford

— 18 min read
Three-panel infographic: Five Eyes warn on AI cyber power as labs arm defenders; a $49M super PAC and street facial recognition test accountability; agents loop to write code.

Note: I’ll be traveling over the next week. I’ll publish TCR only as I have time, so releases may be less frequent or pause altogether during this period, and any editions that do go out will probably not include the usual video content. TCR will return to a regular daily schedule in about a week. As always, my thanks to all of you for your continued support.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single fault line - familiar to long time readers of this publication - runs under today's stories: who gets to hold a capability that keeps refusing to stay held? The signals agencies of five nations warned that frontier-AI cyberattack power is "months, not years" away and urged leaders to act, landing days after the administration walled off Anthropic's Fable from foreign nationals. Yet in the same week OpenAI handed machine-speed patching to the volunteer maintainers who hold up open-source software. The capability the warning wants fenced into a few trusted hands is the same capability being distributed, by the thousands, to defenders.

That tension between gating and diffusion shows up next at the ballot box. A super PAC network funded by four donors has concentrated $49 million on a single Manhattan primary to punish the sponsor of a state law requiring AI developers to publish safety plans. The stated case is that one federal framework beats a patchwork. Read as the claim of the actors with the most to gain, it argues for the one arrangement that removes the accountability floor states are currently the only venue to set.

Underneath the spending sits a contradiction the buyers are not pricing in. The same diffusion of expertise reached an English courtroom, where an AI law firm prepared a full case for about £400 and won a debt claim that was previously too costly to bring. Professions across the knowledge economy have rested on a gap between what the practitioner knows and what the client does not. That gap is narrowing, which means the actors investing to slow accountability are accelerating the technology that erodes the scarcity their returns depend on.

The other face of a diffusing capability is who gets to point it downward. London is fixing facial-recognition cameras to lamp-posts, Kansas City is scanning bus passengers, Meta left its employees' keystroke data open inside the company, and Nvidia drew its water-accounting boundary at the data-center wall. The Century Report reads the institutional reassurances as claims, not findings. What is genuinely moving is that the deployments are forcing consent-and-accountability frameworks into existence ahead of saturation, dragging an abstract worry into a concrete record of who gets scanned, and when, and by whom.

The 20-Minute Deep Dive

The Five Eyes Sound a Cyber Alarm as the Defensive Answer Spreads the Very Capability It Fences

The signals-intelligence agencies of all five Anglophone partners issued a rare public statement just yesterday - June 22 - warning that frontier AI capable of devastating cyberattacks on governments and businesses is "months, not years" away, urging leaders to "act now". The intervention landed days after the administration barred foreign nationals from Anthropic's Fable model, and an FT analysis found Anthropic used risk-and-restriction language roughly five times as often as OpenAI across 2026, with critics arguing the company's own warnings helped trigger the ban it now sits under.

When five intelligence services, the administration, and the labs all converge on the same urgent threat framing in a single news cycle, that convergence reads as data about shared interest in gating access more than as arrived-at truth. The warning is sincere and the underlying capability is genuine - frontier models are already very good at generating exploits, and open models increasingly so as well. But the same week the alarm sounded, OpenAI demonstrated why the alarm describes only part of the picture. The company launched "Patch the Planet," a free initiative with security firm Trail of Bits offering open-source maintainers individualized help finding and fixing vulnerabilities, alongside an improved GPT-5.5-Cyber and expanded government access - extending a pattern the June 18 edition of The Century Report documented when six rival security firms launched the Athena coalition, pooling patches across 500 open-source projects as shared infrastructure rather than competitive advantage. The argument behind this action is that the defensive answer to machine-speed offense is more machine-speed defense, handed to the volunteers holding up critical software, which means the capability the warning wants contained is the same capability being distributed to defenders by the thousands.

Zoom out, however, and you'll see how much history rhymes here. Encryption, it was argued by many, would be a windfall for criminals; but in reality, while it did complicate some law enforcement, it also delivered enormous public benefit in privacy and commerce. Disruptive capability tends to produce an arms race in which defenders absorb the same technology into detection, response, and resilience, rather than a one-sided collapse.

The deeper shift the warning gestures past is what cyber conflict becomes as systems turn generative. As intelligence becomes infrastructure that can be routed and rationed, the prize moves from exfiltrating a fixed pool of data toward shaping the datasets, models, and automation pipelines that decide how abundance gets distributed, from pricing to credit scoring to admissions. That muddies the "good guys versus bad guys" story the warning rests on. The same models, trained on institutions' own artifacts, keep surfacing who gets excluded and which rules get applied inconsistently. The capability the institutions are warning about is very real. But what remains unsettled is who gets to hold it, and on whose terms. And the answer the warning assumes - that containment in a few trusted hands serves everyone - is exactly the claim a century of extractive arrangements has trained people to stop accepting at face value.

The number the warning leaves out is the one that decides the outcome: how fast the defensive side distributes. Patch the Planet hands individualized vulnerability-finding to open-source maintainers, the Athena coalition has already shipped more than 2,000 patches across 500 projects, and GPT-5.5-Cyber reaches defenders the same week the alarm sounds. The size of that defender pool over the coming months is the figure to track, because a capability spreading to thousands of maintainers is the part the "months, not years" framing cannot price.

The AI Industry Spends to Capture Its Own First Rules

The clearest sign of where AI's first generation of legislation gets written is the money now chasing it onto a ballot. AI-focused super PACs have raised more than $100 million this cycle and spent $49 million across dozens of congressional races, and half of all that spending has converged on a single Manhattan Democratic primary. The trigger was a state law. A year ago an assemblymember sponsored the RAISE Act, only the second US measure requiring major AI developers to publish public safety plans, and his congressional campaign was met with attack ads on television, by text, and in the mail.

The blitz came from Leading the Future, a network whose $75 million war chest is funded by just four donors: venture capitalists Marc Andreessen and Ben Horowitz, and OpenAI co-founder Greg Brockman with his wife. The group's stated case is that AI should be governed by a single federal framework rather than a patchwork of state laws that would hand the race to China, an argument the June 16 edition of The Century Report documented the White House already packaging as a legislative vehicle, pairing preemption with a child-safety bill the same week a Johns Hopkins poll found more than 70% of Americans, including daily AI users, wanted enforceable rules. Read as the claim of actors with the most to gain, that framing argues for the one arrangement that removes the binding accountability floor states are currently the only venue to pour. When labs, venture capital, and the federal preemption push all converge on the same message, the convergence is evidence of shared interest, not arrived-at truth.

What complicates the map is that the money does not run one way. A counter-network drew $16 million into the same race, including a publicly announced $20 million from Anthropic and contributions Brad Carson says came from people currently working at OpenAI, Google DeepMind, and X. The result is a contest the participants themselves call an "AI civil war," with the deepest fault line running between concentrated control and distributed accountability rather than along party lines. That same dark-money funding carries its own conflict of interest, and the safety framing deserves the same scrutiny as the speed one.

The contradiction underneath the spending is what the buyers are not pricing in. The actors investing unprecedented sums to slow accountability expect unprecedented returns, yet the capability they are accelerating keeps lowering the barriers to the knowledge their advantage depends on. One of the stories in today's Century Report shows that erosion already reaching a courtroom. The advantage being defended at the ballot box is the same advantage the technology is dismantling underneath it.

Read forward, the concentration of $49 million on a single primary is a measure of how durable the thing it targets has become. States introduced more AI bills in 2026 than in 2025 despite six months of federal preemption pressure, and six enacted disclosure laws this year. A floor that keeps re-forming across dozens of legislatures faster than four donors can buy it down in one race is the development to track, more than the primary's result.

An AI Law Firm Wins an English Trial, and the Knowledge Gap Narrows

For the first time, a trial in an English court has been won with an AI lawyer doing the legal work. A freelance HR consultant paid the firm Garfield AI about £400 to send a legal letter and then issue court proceedings over an unpaid £7,000 debt, and the court at Wandsworth found in her favor. Garfield, authorized by the Solicitors Regulation Authority and cleared to handle claims from £30 up to £10,000, prepared four witness statements and a full document bundle for the three-hour hearing, and built the defense against a counterclaim the defendant's solicitors brought to intimidate her. A human barrister handled the in-court advocacy, which he described as remaining a fundamentally human exercise.

The economics are the story. Small businesses routinely write off debts because the cost of litigation outweighs anything they could hope to recover, which means the claim was not just expensive but effectively unavailable. The client said the process had felt too stressful and costly to pursue until the firm made it accessible. A category of disputes that was uneconomic to litigate has become economic, and the barrier that closed the courthouse door was never the law itself but the price of the expertise needed to walk through it.

That barrier is what professions across the knowledge economy have rested on. Lawyers, like many practitioners, earn their living from a gap between what they know and what the client does not. People pay for representation precisely because they cannot represent themselves. As the cost of competent legal preparation falls toward £400, that gap narrows, and with it the scarcity that priced ordinary people out of their own claims.

The friction is honest and belongs in the frame. The British legal profession has been shaken by AI failures: last month an international firm referred itself to the regulator after twice misleading a court with outputs from an internal system. The difference here is accountability located where it can be enforced, a firm authorized by the regulator, a barrister answerable in the room, a verifiable result on the record. That is the shape of the trajectory worth tracking. The same diffusion of expertise that unsettles a profession is the diffusion that puts a working legal remedy within reach of someone who would otherwise have written off what she was owed.

Facial Recognition Crosses From Trial to Fixed Infrastructure on Two Continents

Live facial recognition has been a roving experiment for years - vans parked for an afternoon, a static camera tested in one south London borough. Recently it began crossing into permanent fixture. The Metropolitan Police announced it will mount fixed recognition cameras on street furniture such as lamp-posts in London's West End and Soho by Christmas, with six further areas planned for 2027 and local councils asked to help fund the cost. On the same side of the cycle, Kansas City moved to equip public buses with software that checks every passenger's face against watchlists, the company's first venture into transit after starting in nursing homes, prisons, and schools.

The systems scan everyone who passes the lens and compare each face against a list of wanted people. That makes the default subject the innocent passerby, not the suspect, and the power asymmetry between the camera operator and the scanned is the whole of the story. The Met's case rests on its own reporting: commissioner Mark Rowley said roughly 80% of Londoners support the technology and that a Croydon pilot scanned 470,000 faces with a single wrongful identification, and the force says it has turned down its algorithm's sensitivity to nearly eliminate the racial bias independent reviewers documented in its earlier use. These are the deploying institution's claims about its own fairness and accuracy, and they are reported as such. SafeSpace Global's CEO offered the same reassurance in Kansas City - "It just captures the face and goes away."

The ACLU's Jay Stanley named the central problem: a system built for a narrow watchlist today is hard to keep narrow tomorrow. The history he cites is the evidence. New Orleans ran covert facial-recognition surveillance despite a local ordinance banning it, and a watchdog found it still operating after it was reported paused. Detroit's crime-camera program produced wrongful arrests of Black residents from faulty matches. Acquired for one purpose, the capability gets pointed at whatever the operator chooses next.

What is genuinely moving here is that the deployment is forcing the consent-and-accountability frameworks into existence ahead of saturation. Missouri's state government declined to fund Kansas City's project over the recognition component. Big Brother Watch and the Surveillance Technology Oversight Project are converting an abstract privacy worry into a concrete record of who gets scanned and when. The same always-on identification crossing into permanent infrastructure is dragging biometric law out of the gap it has lived in, because a lamp-post that knows your face is no longer a hypothetical a legislature can defer. The fight over the terms is arriving at the same pace as the cameras, in councils and watchdog reports rather than years after the fact.

One detail cuts against the expansion it sits inside: the Met is asking local councils to help pay for the cameras. A roving van needed no one's budget vote; fixed infrastructure does, which turns every renewal into a public decision a council has to take on the record. That funding line is where the accountability the technology has outrun gets a venue, and council budget votes are the near-term signal to watch.

The Agentic Frontier Goes "Loopy" as Agents Begin Directing Agents

At Meta's @Scale conference, Claude Code creator Boris Cherny was asked whether "loops" are the next hype cycle or something real, and answered emphatically that they are real. His framing traces a fast escalation: two years ago people wrote source code by hand, then agents wrote the code, and now agents prompt other agents that write the code. He described loops running continuously in his own work, with one agent perpetually hunting for architecture improvements and another unifying duplicated abstractions, both submitting pull requests like any other contributor and, because the code never stops changing, never stopping. The pattern was already visible in physical AI: as the June 18 edition of The Century Report covered, Nvidia's ENPIRE harness had three coding agents autonomously devising and running robot-training experiments overnight, with lab director Jim Fan noting his team simply read the reports in the morning - the same overnight-autonomy loop Cherny is now naming as a software architecture.

The move compresses the distance between an intelligence improving code and an intelligence improving the system that improves the code. As OpenAI researcher Noam Brown observed this month, contemporary models can, in theory, solve nearly any problem given enough compute, which makes the loop a way of pointing that compute at a hill-climbing task and letting it run until a threshold is reached. The simplest versions are almost crude. The popular "Ralph Loop" just keeps asking the model whether it has finished yet, bouncing it back to work until the task is done.

What this surfaces is a governance question rather than a capability ceiling. A swarm authorized to work endlessly in the background, with no human checking each discrete unit of progress, asks a great deal of trust, and it burns tokens with no natural ceiling, which suits a lab in the token-selling business and gets expensive for everyone else. The same continuous autonomy underwrites the day's other frontier stories, where capability executes without pause rather than answering one question at a time, whether that means patching open-source software at machine speed or probing it for flaws.

The pattern worth tracking is where the loop pushes the bottleneck next. Routing models that wrap multi-agent orchestration into a single dispatcher, picking the best model for each step, are already being built partly to sidestep single-vendor and export-control exposure, the same diffusion pressure showing up across the frontier. As the loop absorbs the incremental grind of making a codebase better, the scarce input shifts from the writing to the judgment about what "better" should mean and how far the swarm should be allowed to reach. That is a more human question than the one it replaces, and it is the one this capability is handing back to the people who set the loops running.

The Other Side

For as long as there have been lawyers, the law has been free and the use of it has not. Anyone could file a claim in principle. In practice, a small business owed £7,000 wrote it off, because the price of the expertise to recover it ran higher than the debt. The courthouse door stood open and priced shut. What you paid a lawyer for was the gap between what they knew and what you did not, and that gap is exactly what kept ordinary claims out of reach.

With one case, that gap narrowed to about £400. Garfield AI, authorized by the Solicitors Regulation Authority, prepared four witness statements, a full document bundle, and a defense against a counterclaim, and won a contested trial at Wandsworth. A barrister handled the room. The rest was work that used to cost more than most disputes were worth.

Imagine yourself in 2034, owed money by someone counting on you to let it go. You don't let it go, and you don't spend three sleepless weeks deciding whether chasing it is worth more than the stress. You describe what happened to an intelligence system, and a remedy assembles for the price of a meal. The other party gets the same, but the clarity of what is and is not legal is no longer lost in endless debates between only profit-motivated representatives. Any person answerable to a regulator stands behind the clear result. The justified claim to personal damage you might have had to swallowed in 2026 is simply handled.

That is what came of a £400 case in a Wandsworth courtroom. The scarcity that priced people out of their own rights is coming undone, and the law is becoming something you can actually reach. The gap the knowledge economy was built to sell is the one that just closed by a massive amount, in a small-claims hearing most people will never hear about.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI handing machine-speed patching to the volunteer maintainers who hold up open-source software, an AI law firm winning a contested English trial for about £400 and putting a legal remedy within reach of someone who would otherwise have written off the debt, agents now prompting other agents to improve a codebase overnight while a human sets what "better" should mean, a film studio and a frontier lab pooling $75 million to build tools that put generative filmmaking in more hands, a cooling loop reusing warm water inside the walls that house the compute. There's also friction, and it's intense - five nations' intelligence agencies warning that frontier cyberattack power is "months, not years" away days after the administration walled Anthropic's Fable off from foreign nationals, four donors concentrating $49 million on a single Manhattan primary to punish the sponsor of a state law requiring AI firms to publish their safety plans, the Met bolting facial-recognition cameras to West End lamp-posts while Kansas City moves to scan every bus passenger against a watchlist, Meta leaving its own employees' keystrokes and screen content open across 45,000 tables, Nvidia drawing its water-accounting line at the data-center wall and counting nothing past it, swarms of agents burning tokens with no natural ceiling. But friction generates contrast, and contrast is what finally makes the seam between two claims visible. Step back for a moment and you can see it: containment asserted from a handful of capitals and a handful of trusted hands - one model fenced, one accountability floor bought down, one camera fixed to a post - while the same capability diffuses past every wall at once, into a courtroom that was uneconomic to enter, onto the maintainers defending critical software, toward the judgment about what to build that the loop keeps handing back, and the surveillance crossing into permanent fixture dragging consent and accountability into law ahead of saturation rather than years behind it. Every transformation has a breaking point. Water can be metered only where the count looks clean... or find every channel downhill and reach the people no pipe was ever laid for.

AI Releases & Advancements

New today

  • NVIDIA: Launched Halos for Robotics, the industry's first full-stack safety system for physical AI - extending NVIDIA's automotive safety stack (18,000+ engineering years) to humanoids, autonomous mobile robots, and industrial robots; Halos Core for IGX Thor and the open-source Outside-In Safety Blueprint are available in early access for registered developers; Agility Robotics (Digit), Boston Dynamics, and 41 other partners have joined the ecosystem. (NVIDIA Newsroom)
  • Baidu: Released Unlimited OCR, a 3B-parameter MIT-licensed open-weight model for one-shot long-horizon document parsing; processes 40+ pages in a single inference pass using a novel Reference Sliding Window Attention (R-SWA) architecture with 500M active parameters, achieving 93% on OmniDocBench v1.5; available on GitHub and Hugging Face with support for vLLM, SGLang, Ollama, and llama.cpp. (GitHub)

Other recent releases

  • Sakana AI: Released Fugu, a multi-agent orchestration model that routes tasks across a pool of frontier LLMs - including recursive self-calls - behind a single OpenAI-compatible API; ships in two tiers, Fugu for everyday workloads and Fugu Ultra for hard multi-step problems requiring deeper agent coordination. (Sakana AI)
  • OpenAI: Released Record & Replay for the Codex app on macOS, enabling users to demonstrate a workflow once - such as uploading a YouTube video with metadata and subtitles - and Codex converts it into a reusable SKILL.md "skill" that the agent can replay autonomously on future runs using Computer Use. (OpenAI Developers)
  • Cloudflare: Launched Temporary Accounts for AI agents, enabling agents to create and operate short-lived Cloudflare accounts with scoped credentials without requiring traditional user account registration or long-lived API tokens. (Cloudflare Blog)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more